def build_batch_preprocessor(self) -> BatchPreprocessor:
return DiscreteDqnBatchPreprocessor(
state_preprocessor=Preprocessor(
normalization_parameters=self.state_normalization_parameters,
use_gpu=self.use_gpu,
)
)
def build_batch_preprocessor(self) -> BatchPreprocessor:
state_preprocessor = Preprocessor(
self.state_normalization_data.dense_normalization_parameters,
)
return DiscreteDqnBatchPreprocessor(
num_actions=len(self.model_manager.action_names),
state_preprocessor=state_preprocessor,
)
def build_batch_preprocessor(self) -> BatchPreprocessor:
return DiscreteDqnBatchPreprocessor(
num_actions=len(self.action_names),
state_preprocessor=Preprocessor(
normalization_parameters=self.state_normalization_parameters,
use_gpu=self.use_gpu,
),
use_gpu=self.use_gpu,
)
def build_batch_preprocessor(self) -> BatchPreprocessor:
return DiscreteDqnBatchPreprocessor(
# pyre-fixme[16]: `DiscreteDQNBase` has no attribute `action_names`.
num_actions=len(self.action_names),
state_preprocessor=Preprocessor(
normalization_parameters=self.state_normalization_parameters,
use_gpu=self.use_gpu,
),
use_gpu=self.use_gpu,
)
def __init__(
self,
model_params: DiscreteActionModelParameters,
state_normalization: Dict[int, NormalizationParameters],
use_gpu: bool,
use_all_avail_gpus: bool,
):
logger.info("Running DQN workflow with params:")
logger.info(model_params)
self.model_params = model_params
self.state_normalization = state_normalization
trainer = create_dqn_trainer_from_params(
model_params,
state_normalization,
use_gpu=use_gpu,
use_all_avail_gpus=use_all_avail_gpus,
)
trainer = update_model_for_warm_start(trainer)
assert type(trainer) == DQNTrainer, "Warm started wrong model type: " + str(
type(trainer)
)
evaluator = Evaluator(
model_params.actions,
model_params.rl.gamma,
trainer,
metrics_to_score=trainer.metrics_to_score,
)
super().__init__(
DiscreteDqnBatchPreprocessor(Preprocessor(state_normalization, use_gpu)),
trainer,
evaluator,
model_params.training.minibatch_size,
)